Rotated FCOS vs Oriented R-CNN on macOS

In June we ran Oriented R-CNN on a MacBook with --device mps — no CUDA toolchain, one CLI command, rotated boxes on a real aerial tile. v0.2.0 added a fourth detector family: Rotated FCOS, an anchor-free single-stage model with a decoded-rIoU 3× Hub checkpoint.

This post puts both on the same Mac and the same images. Same canvas, same NMS, same score floor — Apple M1 Max, PyTorch MPS, Hub 3× weights. We also pull the training wall times from the published Hub runs (same NVIDIA L4 recipe). The question is practical: does the new one-stage model feel good enough on a laptop to replace the two-stage demo default?

Short answer: yes for most scenes, with one caveat about score thresholds. FCOS is faster at inference and about 5.7× faster to train, slightly more accurate on the published DOTA eval-val protocol, and visually competitive — but its scores are less peaked, so the old --score-thr 0.7 from the Oriented R-CNN walkthrough will silently drop half the boxes.

What we compare Link to heading

Both are ResNet-50 + FPN, DOTA le90, train+val pretrain / val eval. The current Hub asset is the refreshed decoded-rIoU run at 82.32%.

Inference numbers below: Apple M1 Max, 64 GB unified memory, PyTorch 2.13.0, oriented-det 0.2.0, --device mps. Protocol: score ≥ 0.3, merge NMS IoU ≤ 0.1 (FCOS eval NMS; same floor for both). Training wall times come from the Hub run logs on a single NVIDIA L4 (see below).

Quick start Link to heading

From an oriented-det checkout with the usual macOS install (uv pip install torch torchvision then uv pip install -e .):

odet pretrained download rotated_fcos_dota_le90_3x
odet pretrained download oriented_rcnn_dota_le90_3x

odet image-demo demo/demo.jpg hf://rotated_fcos_dota_le90_3x \
  --out-file demo_fcos.png \
  --device mps \
  --score-thr 0.3 \
  --nms-thr 0.1

Swap the slug for Oriented R-CNN. Keep --nms-thr 0.1 unless you have a reason to match a two-stage production config (0.3). And prefer --score-thr 0.3 (or 0.25) for FCOS — more on that below.

Bus lot: same scene as the June demo Link to heading

demo/demo.jpg is the 1024×1024 DOTA tile from the macOS walkthrough — diagonal buses and trucks, the scene where axis-aligned boxes look silly.

Counts at score ≥ 0.3 are nearly identical (102 vs 100). Box geometry looks right on both: headings follow the chevron parking, large-vehicle / small-vehicle labels match. The visible difference is score calibration — Oriented R-CNN piles many boxes near 1.00; FCOS spreads them across roughly 0.4–0.9. That is expected for a one-stage sigmoid head versus a two-stage ROI classifier, but it changes which CLI threshold you want.

Score thresholds: do not copy --score-thr 0.7 onto FCOS Link to heading

The June post used --score-thr 0.7 for a clean Oriented R-CNN overlay. On that same checkpoint family it still works. On FCOS it does not:

At 0.3 the two models agree. At 0.7, FCOS keeps fewer than half the boxes while Oriented R-CNN barely flinches. For FCOS demos, start at 0.25–0.30 (the train-val / F1-maximizing region from the v0.2.0 report).

Latency on MPS Link to heading

Timed after warmup; mean of 5 single-forward runs (or 3 for the tiled harbor). Canvas 1024×1024, ORIENTED_DET auto window batch 32 on MPS.

On sparse tiles the one-stage head is almost 2× faster. On the dense bus lot the gap shrinks because final oriented NMS (Python, AABB-prefiltered) scales with detection count — both emit ~100 boxes, so NMS dominates. The harbor tile (six overlapping windows, ~300 ships) still favors FCOS by about 40%.

Neither model needs custom CUDA kernels. MPS just works.

Training wall time (same L4 recipe) Link to heading

The Hub 3× checkpoints were trained on the same DOTA tile recipe: 13,691 train+val tiles, batch size 2, 36 epochs, milestones at 24 / 33, single NVIDIA L4. Timings are from the published sidecar logs (oriented_rcnn_r50_fpn_dota_le90_3x-68957f98.log, rotated_fcos_r50_fpn_dota_le90_3x-6e383331.log).

Rotated FCOS finishes the 3× schedule in under a day — about 5.7× less wall clock than Oriented R-CNN on the same GPU and data. Train-only epochs (no periodic mAP) sit around 31–33 minutes for FCOS versus a few hours for Oriented R-CNN; mAP-every-4 epochs stretch both (FCOS peaks near 57 minutes when GPU-sampled val mAP runs).

That gap is architectural: Oriented R-CNN pays for an RPN plus oriented RoIAlign on every proposal every step. FCOS is a single dense head over P3–P7. The July ProbIoU post already showed Rotated Faster R-CNN (~58 min/epoch) beating Oriented R-CNN on training cost; FCOS lands in a similar per-epoch band while staying one-stage and anchor-free.

If you are iterating recipes on a single L4, that is the difference between waiting overnight and waiting almost a week for a 3× run.

Harbor: sliding-window ships Link to heading

Same large.jpg as the v0.1.1 harbor demo (1299×1904). The DOTA recipe tiles oversized rasters; here that is six 1024 windows with 200 px overlap.

odet image-demo demo/large.jpg hf://rotated_fcos_dota_le90_3x \
  --out-file large_fcos.png \
  --device mps \
  --score-thr 0.3 \
  --nms-thr 0.1

Ship counts match almost exactly. Visually both cover the moored rows; FCOS finishes in under two seconds on this machine.

Planes and storage tanks Link to heading

Two more DOTA val tiles for class variety — airport apron and tank farm — same thresholds.

Both find the nine aircraft. FCOS scores sit in the 0.8s; Oriented R-CNN saturates at 1.00. Parking-lot vehicles are comparable.

Compact circular tanks are a published FCOS strength on eval-val (ahead of Faster R-CNN on that class in the v0.2.0 note). On this tile both models land the visible tanks cleanly.

When to pick which Link to heading

For laptop demos after v0.2, Rotated FCOS 3× is the better default than Oriented R-CNN 1×/3×: higher published mAP50, fewer parameters, faster MPS inference, much cheaper 3× training, and no RPN. Keep score thresholds in the 0.25–0.30 band and NMS at 0.1.

Commands (copy-paste) Link to heading

# Prefetch
odet pretrained download rotated_fcos_dota_le90_3x
odet pretrained download oriented_rcnn_dota_le90_3x

# FCOS on the June bus-lot tile
odet image-demo demo/demo.jpg hf://rotated_fcos_dota_le90_3x \
  --out-file demo_fcos.png --device mps --score-thr 0.3 --nms-thr 0.1

# Side-by-side Oriented R-CNN
odet image-demo demo/demo.jpg hf://oriented_rcnn_dota_le90_3x \
  --out-file demo_orcnn.png --device mps --score-thr 0.3 --nms-thr 0.1

# Harbor (sliding windows)
odet image-demo demo/large.jpg hf://rotated_fcos_dota_le90_3x \
  --out-file large_fcos.png --device mps --score-thr 0.3 --nms-thr 0.1

References Link to heading

• oriented-det on GitHub
• Pretrained zoo — rotated_fcos_dota_le90_3x, oriented_rcnn_dota_le90_3x (sidecar .log files hold the train timing summaries)
• Rotated FCOS recipes
• Previous: Oriented-Det v0.2.0 · macOS Oriented R-CNN walkthrough · Faster R-CNN training-cost context

Written on September 2, 2026 by Jeff Faudi. Link to heading